RNN-MBP

Deep Recurrent Neural Network with Multi-scale Bi-directional Propagation for Video Deblurring (AAAI-2022)
by Chao Zhu, Hang Dong, Jinshan Pan, Boyang Liang, Yuhao Huang, Lean Fu, and Fei Wang

[Paper] [Supp]

Results

Results on GOPRO

Results on DVD

Results on RBVD

Prerequisites

• Python 3.6
• PyTorch 1.8
• opencv-python
• scikit-image
• lmdb
• thop
• tqdm
• tensorboard

Real-world Bluryy Video Dataset (RBVD)

We have collected a new RBVD dataset with more scenes and perfect alignment, using the proposed Digital Video Acquisition System.

Training

Please download and unzip the dataset file for each benchmark.

• GOPRO
• DVD
• RBVD

Then, specify the \<path> (para.data_root) where you put the dataset file and the corresponding dataset configurations in the command (e.g. para.dataset=gopro or gopro_ds_lmdb).

The default training process requires at least 4 NVIDIA Tesla V100 32Gb GPUs.

The training command is shown below:

python main.py --data_root <path> --dataset gopro_ds_lmdb  --num_gpus 4 --batch_size 4  --patch_size [256, 256]  --end_epoch 500

Testing

Please download checkpoints and unzip it under the Source directory.

Example command to run a pre-trained model:

python test.py --data_root <path> --dataset gopro_ds_lmdb  --test_only --test_checkpoint <path>  --model RNN-MBP 

Citing

If you use any part of our code, or RNN-MBP and RBVD are useful for your research, please consider citing:

@inproceedings{chao2022,
  title={Deep Recurrent Neural Network with Multi-scale Bi-directional Propagation for Video Deblurring},
  author={Chao, Zhu and Hang, Dong and Jinshan, Pan and Boyang, Liang and Yuhao, Huang and Lean, Fu and Fei, Wang},
  booktitle={AAAI},
  year={2022},
}